Research & Education

DR Tulu is an open-source Deep Research (DR) model designed for tackling long-form research tasks. The DR Tulu-8B model has demonstrated performance comparable to OpenAI DR on long-form DR benchmarks. This repository provides the official code for DR Tulu, including an agent library with a MCP-based tool backend, high-concurrency async request management, and a flexible prompting interface for developing and training deep research agents. It also includes RL training code based on Open-Instruct and SFT training code based on LLaMA-Factory, allowing for supervised fine-tuning and reinforcement learning with GRPO and evolving rubrics. An interactive CLI demo is available for users to experiment with DR Tulu-8B.

EconML is a Python package developed by Microsoft Research as part of the ALICE (Automated Learning and Intelligence for Causation and Economics) project. It provides a toolkit for estimating heterogeneous treatment effects from observational data, integrating advanced machine learning techniques with econometrics. The package is designed to measure the causal effect of treatment variables on an outcome, controlling for various features, and how this effect varies. It supports methods like Double Machine Learning, Causal Forests, Orthogonal Random Forests, and Meta-Learners, offering flexibility in modeling effect heterogeneity while preserving causal interpretation and providing confidence intervals. EconML is built on standard Python packages for Machine Learning and Data Analysis, making it accessible for data scientists and researchers.

EntitySeg is an open-source toolbox designed for advanced image segmentation tasks, focusing on open-world and high-quality segmentation. It consolidates several cutting-edge algorithms developed by the qqlu group, including Open-World Entity Segmentation (TPAMI2022), High Quality Segmentation for Ultra High-resolution Images (CVPR2022), CA-SSL: Class-Agnostic Semi-Supervised Learning (ECCV2022), and High-Quality Entity Segmentation (ICCV2023 Oral). The toolbox is built using Python and PyTorch, making it accessible for researchers and developers in the computer vision domain. It aims to provide a unified platform for various image segmentation challenges, with future plans to merge all projects for enhanced interoperability and support.

The MT-Reading-List is a comprehensive resource for researchers and students interested in machine translation. Maintained by the Tsinghua Natural Language Processing Group, it offers a curated collection of papers spanning the evolution of the field, from statistical machine translation (SMT) to neural machine translation (NMT). While it prioritizes contemporary NMT papers, it also acknowledges historical context, referencing older papers. The list is continuously updated and categorized, covering various sub-topics like model architectures, attention mechanisms, low-resource translation, multilingual MT, robustness, interpretability, and efficiency. It also includes sections for "10 Must Reads" and WMT winners, making it a valuable starting point for anyone delving into machine translation research.

efficient-gnns is a comprehensive repository offering code and resources for developing scalable and efficient Graph Neural Networks (GNNs). It specifically focuses on knowledge distillation techniques, including novel approaches like Graph Contrastive Representation Distillation, to create resource-efficient GNNs. The repository benchmarks various distillation methods, such as Local Structure Preserving loss and Global Structure Preserving loss, alongside baselines like Logit-based KD. It supports research on large-scale, real-world graph datasets for tasks like graph classification on MOLHIV and node classification on ARXIV and MAG, providing installation and usage instructions for researchers and developers in the field.

Federated is a collection of Google research projects dedicated to advancing Federated Learning and Federated Analytics. Federated learning enables the training of a shared global model across numerous participating clients while ensuring their training data remains local. Federated analytics, on the other hand, focuses on applying data science methods to analyze raw data stored directly on users’ devices. Many projects within this repository leverage TensorFlow Federated (TFF), an open-source framework designed for machine learning and other computations on decentralized data. The repository serves primarily for reproducing experimental results from related papers, with each project intended as an independent unit rather than a reusable framework.

Fewshot_Detection is an open-source implementation of the paper "Few-shot Object Detection via Feature Reweighting," designed for researchers and developers working with computer vision. This tool addresses the challenge of detecting novel objects with limited training data by employing a meta feature learner and a reweighting module within a one-stage detection architecture. It is built upon `pytorch-yolo2` and developed with Python 2.7 and PyTorch 0.3.1. The system extracts meta features generalizable to novel object classes and transforms support examples into reweighting vectors, enhancing detection capabilities. The entire process, including a carefully designed loss function, is trained end-to-end based on an episodic few-shot learning scheme. It demonstrates significant performance improvements over established baselines on multiple datasets and settings.

PiML-Toolbox (Python Interpretable Machine Learning) is a comprehensive Python toolbox designed for the development and diagnostics of interpretable machine learning models. It offers both low-code interfaces and high-code APIs, supporting a growing list of inherently interpretable ML models such as GLM, GAM, Tree, FIGS, XGB1, XGB2, EBM, GAMI-Net, and ReLU-DNN. The toolbox facilitates various outcome testing, including accuracy, explainability (PFI, PDP, ALE, LIME, SHAP), fairness, weak spot identification, overfitting detection, reliability assessment, robustness, and resilience evaluation. PiML-Toolbox aims to empower model developers and validators with tools for transparent, interpretable, and robust machine learning, particularly in high-stakes regulatory settings.

GNN-Recommender-Systems is a valuable resource for researchers and developers focused on recommendation algorithms utilizing Graph Neural Networks (GNNs). This index compiles a wide array of GNN-based recommendation algorithms, categorized by different recommendation stages (matching, ranking, re-ranking), scenarios (social, sequential, session, bundle, cross-domain), and objectives (multi-behavior, diversity, explainability, fairness). Each entry includes the algorithm's name, associated paper, publication venue, year, and often a link to its code implementation. The project is based on a survey paper published in ACM Transactions on Recommender Systems, offering a structured overview of the field.

The Connected Ideas Project, by Alexander Titus, is a Substack publication dedicated to exploring the intricate connections between technology, policy, people, and ideas. It delves into the impact of emerging technologies such as AI, biotechnology, fusion, and quantum on our lives, often incorporating elements of science fiction. With thousands of subscribers, this platform offers insights and analysis for those interested in the evolving landscape of innovation and its societal implications. The project aims to provide a comprehensive understanding of how these diverse fields intersect and influence each other.

REP, or Reproducible Experiment Platform, is an ipython-based environment designed for conducting data-driven research with an emphasis on consistency and reproducibility. It provides a unified Python wrapper for several machine learning libraries, including Sklearn, XGBoost, and Theanets, allowing users to work with a consistent interface. Key features include parallel training of classifiers on clusters, classification/regression reports with interactive plots, and smart grid-search algorithms with parallel execution. REP also supports research versioning using Git and offers pluggable quality metrics for classification. It aims to extend scikit-learn by providing a better user experience and tools for meta-algorithm design, making it a valuable resource for data scientists and researchers.

pointnet.pytorch offers a PyTorch implementation of the PointNet deep learning model, specifically designed for 3D classification and segmentation using point sets. This open-source tool facilitates research and development in 3D data processing, providing a robust and tested framework compatible with PyTorch 1.0. It includes functionalities for downloading and preparing datasets, training classification and segmentation models, and visualizing results. The repository details performance metrics on datasets like ModelNet40 and ShapeNet, allowing users to compare against original implementations. It's a valuable resource for developers and researchers working with 3D point cloud data.

Counsel Stack offers an enterprise-grade legal citation verification API designed for legal professionals. It helps detect and correct over 40 categories of legal errors, including hallucinated citations, technical inaccuracies, fabricated holdings, and overturned cases. The platform is built to withstand legal scrutiny, ensuring attorneys can certify legal arguments under Rule 11 with confidence. Counsel Stack also provides a Research API to answer complex legal questions, outperforming generalist AI and lawyer baselines in independent benchmarks. It includes comprehensive federal law coverage, over 99% of precedential case law, and an expanding collection of state legal sources, accessible via API or local deployment. This tool is efficient and scalable, processing over 100 cite checks per minute at an average cost of $0.0085 per check.

R1-V is an open-source project focused on enhancing the super generalization ability of Vision Language Models (VLM) with minimal computational cost. It aims to improve the perception and reasoning capabilities of VLMs through reinforcement learning. The project provides new VLM-RL environments, a comprehensive training codebase, and research papers. R1-V supports various models like Qwen2-VL and Qwen2.5-VL, and offers training datasets for tasks such as item counting and geometry reasoning. It also includes evaluation scripts for benchmarks like SuperClevr and GEOQA, making it a valuable resource for researchers and developers in the VLM domain.

Handwriting-generation is an open-source project that provides an implementation of handwriting generation using recurrent neural networks in TensorFlow. Based on Alex Graves' research paper (https://arxiv.org/abs/1308.0850), this tool enables users to download datasets, preprocess them, train their own models, and then generate handwriting. It offers options to control generation parameters like bias for clarity, visualize the writing process with animation, and even select different handwriting styles. This makes it a valuable resource for AI researchers and developers interested in exploring and experimenting with handwriting synthesis.

TextGAN-PyTorch is a comprehensive PyTorch framework designed for Generative Adversarial Networks (GANs) based text generation models. It supports both general and category-specific text generation, making it a versatile tool for researchers and developers. The framework serves as a benchmarking platform, facilitating the evaluation and comparison of various GAN-based text generation models. It is particularly beneficial for those familiar with PyTorch, enabling them to quickly engage with the text generation field. The repository includes implementations of several prominent models like SeqGAN, LeakGAN, and RelGAN, along with detailed instructions for setup and usage, including real data experiments and visualization tools.

ViT-pytorch offers a PyTorch reimplementation of the Vision Transformer (ViT) model, based on the paper 'An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale'. This tool allows users to leverage the power of Transformers for image recognition, demonstrating that applying them directly to image patches and pre-training on large datasets yields state-of-the-art results. It includes various pre-trained models like ViT-B_16, R50+ViT-B_16, and ViT-L_32, which can be downloaded and used for training. The repository provides scripts for training models on datasets like CIFAR-10 and CIFAR-100, with options for mixed precision training and gradient accumulation. Additionally, it supports visualization of attention maps, offering insights into how the model processes images.

Unholy.ai is an AI-powered tool designed to scan song lyrics and audio for potentially offensive or controversial content. It helps identify themes such as erotica, blasphemy, adultery, and other explicit content, providing a detailed analysis for content management. This enables platforms and users to maintain specific content standards and ensure compliance with ethical guidelines. The tool aims to help users detect any "unholiness" in the music they listen to, offering a solution for those who need to filter or understand the thematic content of songs.

ViTPose is an official PyTorch implementation for human pose estimation, based on the NeurIPS'22 paper "ViTPose: Simple Vision Transformer Baselines for Human Pose Estimation" and the TPAMI'23 paper "ViTPose++: Vision Transformer for Generic Body Pose Estimation." This tool achieves impressive accuracy, including 81.1 AP on the MS COCO Keypoint test-dev set. It supports both single-task and multi-task training, covering human, animal, and whole-body pose estimation. ViTPose provides pre-trained models, detailed configurations, and a web demo integrated into Huggingface Spaces for easy experimentation with videos and images. It's built on PyTorch and utilizes mmcv, making it a robust solution for researchers and developers in computer vision.

VLM2Vec is an open-source project from TIGER-AI-Lab, providing a unified framework for training and evaluating powerful multimodal embeddings across diverse visual formats, including images, videos, and visual documents. It introduces MMEB-V2, a comprehensive benchmark with 78 tasks designed to systematically evaluate embedding models across these modalities. VLM2Vec-V2 sets a new state-of-the-art, outperforming strong baselines. The tool supports easy configuration of training and evaluation using YAML files and allows for easy extension with new datasets. It is built on state-of-the-art Vision-Language Models like Qwen2-VL, using instruction-guided contrastive training to produce fixed-dimensional embeddings for various inputs.

Explore Clinical & Biomedical Language Models is an AI tool hosted on Hugging Face Spaces, designed for researchers and developers in the healthcare sector. This platform facilitates the investigation and evaluation of various language models specifically tailored for clinical and biomedical applications. It serves as a valuable resource for understanding the capabilities and performance of these specialized AI models, aiding in research and development efforts within the medical and life sciences fields. The tool aims to provide a centralized location for discovering and interacting with advanced language models relevant to clinical and biomedical data.

NotedSource is an AI-enhanced R&D platform designed to streamline the research and development lifecycle for companies. It leverages AI to connect organizations with the appropriate expertise, supporting projects from initial concept to final results. Key features include Research Compass AI, which distills hundreds of publications into focused insights, and an AI-powered research collaboration platform for efficient project execution. NotedSource also offers Niche Expertise for AI Annotation and Evaluation through Scholar Data Services, helping accelerate AI development with clean, trustworthy data. The platform provides centralized project management, built-in legal and compliance support, and access to a network of over 50,000 contributors across various fields.

Must-read-papers-and-continuous-tracking-on-Graph-Neural-Network-GNN-progress is a comprehensive open-source project hosted on GitHub, dedicated to compiling and continuously updating a list of must-read papers on Graph Neural Networks (GNN). It serves as a valuable resource for researchers and practitioners interested in the rapidly evolving field of GNNs, which are crucial for analyzing graph-structured data in various domains like social networks, bioinformatics, and computer vision. The project tracks significant advancements, including highly cited works and recent publications from top conferences, providing a centralized hub for key academic contributions. It aims to support researchers by offering a curated selection of foundational and cutting-edge literature.

multiagent-particle-envs offers a foundational code environment for researchers to explore multi-agent systems, particularly in the context of mixed cooperative-competitive environments. This tool, developed by OpenAI, provides a simple multi-agent particle world with continuous observation and discrete action spaces, alongside basic simulated physics. It was specifically used in the paper "Multi-Agent Actor-Critic for Mixed Cooperative-Competitive Environments." While the original repository is archived and read-only, a maintained version with numerous fixes, comprehensive documentation, pip installation support, and compatibility with current Python versions is available in PettingZoo. The environment allows for the creation and testing of various scenarios, including cooperative navigation, predator-prey dynamics, and communication tasks.